Research project · First author

Virtual Roads,
Smarter Safety

A physics-aware digital twin for testing autonomous vehicles in mixed traffic before the hard cases reach public roads.

I developed a field-to-simulation workflow that reconstructs real roads, synchronizes traffic and autonomous-vehicle behavior, models tire and vehicle dynamics, and turns each run into measurable safety evidence.

Read the paper Back to research

A georeferenced three-dimensional reconstruction of a real roundabout from aerial LiDAR and map data — A real intersection reconstructed as a simulation-ready 3D environment.

3 enginesSUMO · CARLA · NVIDIA PhysX

8 scenariosRear-end, sideswipe, blind-spot, and intersection cases

59% lower MAEHigh-fidelity TTC versus traditional TTC

The problem

Safety metrics are only as credible as the world behind them.

Crash records tell us what already failed. Public-road testing cannot efficiently cover rare, dangerous events. And conventional time-to-collision calculations often flatten the problem to constant speed and one-dimensional motion.

Those shortcuts can omit the factors that decide whether an AV stops safely: road slope, intersection geometry, tire friction, mass distribution, steering, suspension, and the behavior of surrounding traffic.

The design goal was end-to-end fidelity: realistic roads, realistic traffic, realistic vehicle physics, and a risk measure that connects them.

The system

One synchronized pipeline connects the physical road to the safety metric.

Architecture linking a reconstructed environment, SUMO traffic, CARLA simulation, and NVIDIA PhysX vehicle dynamics — The platform keeps road geometry, traffic state, sensor-rich simulation, and vehicle dynamics synchronized.

Real to digital

Reconstruct the environment

Aerial LiDAR and OSM establish the scene; GPS and inclinometer measurements refine slope, superelevation, and curvature.

Traffic to autonomy

Run mixed traffic

SUMO controls network-scale traffic and signals while CARLA hosts the ego vehicle, sensors, environment, and custom AV logic.

Physics to evidence

Measure what changes risk

PhysX resolves wheel and tire forces; the safety layer evaluates collision risk in the full three-dimensional scene.

Environment reconstruction

Build the road at the fidelity the vehicle will feel.

The platform combines complementary sources instead of asking one dataset to do everything. OSM supplies accessible network structure, aerial LiDAR captures the physical scene, and vehicle-mounted sensors recover roadway attributes that influence dynamics.

Semantic segmentation separates roads, buildings, and vegetation in the point cloud.
RoadRunner converts the georeferenced scene into OpenDRIVE and simulation assets.
GPS and inclinometer data sampled at 5 Hz refine curvature, slope, and cross-slope.
DBSCAN-based junction assignment repairs missing intersection connections.

Field collection: configuring the drone's automated 3D capture and surveying the real road environment.

Watch the real-to-digital step

Capture once. Reuse the environment across AV experiments.

The drone follows a structured scan plan to collect overlapping multi-view imagery. Ground control points anchor the reconstruction to real coordinates so the resulting geometry can be aligned with OSM and exported to the simulation stack.

Drone three-dimensional scan setup interface in the field — Construction workflowFrom field observations to a simulation-ready digital twin

Photogrammetric reconstruction of the surveyed roadway environment — Construction workflowFrom field observations to a simulation-ready digital twin

High-risk scenario testing

Risk is designed into the road, then tested under live traffic.

ASAM OpenDRIVE gives the platform a parametric description of roads, lanes, markings, objects, and signals. That lets us create repeatable safety-critical tests by changing the infrastructure geometry itself—not just the actors moving through it.

Scenario-generation methodSelect · Perturb · Assemble

Start from a reusable road template, modify the safety-critical geometric parameter, and deploy the resulting network into the digital twin.

01Select a template

Choose an OpenDRIVE lane drop, on-ramp, or weaving interface.

02Perturb geometry

Shorten the taper, acceleration lane, or distance between entry and exit.

03Assemble & run

Generate the network, add traffic, and observe vehicle behavior and safety telemetry.

These are geometry-controlled tests: each hazard can be reproduced, varied, and compared while the platform records the same simulation and safety channels.

Physics experiment

The same control input can produce a different safety outcome.

With throttle held at 0.5, the demonstration changes the tire-friction scaling factor from 5.0 to 1.0. The synchronized view exposes the resulting wheel forces, slip behavior, and vehicle spacing—not just the rendered motion.

Use the controls to compare the two runs.

Higher gripTire friction 5.0 · throttle 0.5

Lower gripTire friction 1.0 · throttle 0.5

Safety evaluation

Simulation becomes engineering evidence when the metric tracks the physics.

Prediction and collision checking running inside the mixed-traffic digital twin.

The paper evaluates eight synthetic collision scenarios spanning rear-end, sideswipe, blind-spot, and intersection conflicts. It compares traditional TTC with simulated collision time and a high-fidelity TTC that accounts for multidimensional motion and roadway effects.

0.29 sHigh-fidelity TTC mean absolute error, versus 0.71 s for traditional TTC.

MethodMAERMSEMaximum error

Traditional TTC0.71 s0.97 s2.17 s

High-fidelity TTC0.29 s0.32 s0.60 s

That is a 59% reduction in MAE and a 67% reduction in RMSE in the paper's reported comparison.

My contribution

I worked across the stack—from infrastructure data to AV safety analysis.

Digital-twin construction

Built HD environments from LiDAR point clouds, OpenStreetMap, RoadRunner, and vehicle-sensor data.

Python co-simulation

Integrated CARLA, SUMO, and NVIDIA PhysX into a synchronized mixed-traffic analysis platform.

Safety experiments

Connected controllable scenarios and physics parameters to multidimensional collision-risk evaluation.

Research leadership

Led the first-author paper that brings mapping, transportation simulation, vehicle dynamics, and safety into one system.

Industry relevance

Transportation context is an advantage for autonomous-driving validation.

This work shows how I approach AV systems: the vehicle, road, traffic stream, infrastructure, and safety metric are coupled parts of one engineering problem.

Reproducible rare-event testing
Infrastructure-aware scenario design
Physics sensitivity analysis
Extensible control and sensor evaluation

Publication

H. Zhang, X. Yue, K. Tian, S. Li, K. Wu, Z. Li, D. Lord, and Y. Zhou, “Virtual Roads, Smarter Safety: A Digital Twin Framework for Mixed Autonomous Traffic Safety Analysis,” IEEE Internet of Things Journal.

Paper and citation

Let's connect

Interested in simulation, digital twins, or AV safety validation?

haozhangch9811@gmail.com

Virtual Roads,
Smarter Safety

Safety metrics are only as credible as the world behind them.